ICLR 2026 · Workshop on Advances in Financial AI · Poster
Beyond Vector Search: Hallucination-Free Financial Reasoning with Chunk-Centric Knowledge Graphs
An agentic knowledge graph for institutional finance.
- 94%
- accuracy on financial document QA
- +8%
- above vector-RAG baselines
- 1
- new benchmark: BeigeBench
Vector search finds passages that sound like the question. For financial documents, close enough is how a model ends up confidently citing the wrong number. This work swaps similarity for structure: resolve the entities, connect them through a constrained financial ontology, and only then pull the passage that holds the answer.
- Entity
- Relationship
- Chunk
Every answer can be traced back along that path to its source evidence. I also built BeigeBench, a reasoning benchmark across multiple Beige Book districts, and released the retrieval system as an open-source Python package.
See it at ICLR →
